Dynamic Capability Scoping for Enterprise AI Agents: A Synthetic Dataset and Three-Source Permission Architecture

TL;DR - This paper proposes dynamically granting enterprise AI agents only the permissions needed for each task, reducing persistent over-privilege. It also releases a 600-prompt synthetic dataset for evaluating permission-scoping systems.

• The architecture combines role-based permission ceilings, task-context classification, and policy-defined permission-combination prohibitions.
• It supports enforcement and observe-only modes, with the latter logging anomalous permission requests for misalignment research.
• The dataset labels minimum required access across 15 tool-permission categories using separate prompt-generation and labeling passes.
• Human review showed high agreement (Cohen’s κ of 0.917 before review and 0.967 after); joint policy-dataset iteration reduced ceiling violations from 46 to 3.